Design And Research Of Target Detection System Based On Zynq

Target detection technology has a wide range of applications in people's lives.Traditional target detection technology is not very robust to various targets,and has encountered an insurmountable bottleneck.With the development of deep learning technology,convolutional neural networks have higher accuracy and robustness than traditional target detection technologies.However,as a computationally intensive algorithm,target detection algorithms based on deep learning are available on portable devices.The realization of this is facing huge challenges.In theory,the two-stage detector has higher accuracy,but it needs to consume a lot of computing resources and time.The Yolo series target detection algorithm under the one-stage framework can obtain the predicted bounding box and category probability through the processing of a single convolutional neural network.Its detection speed is faster and the structure is simpler.In particular,Yolo V4 turned out to achieve relatively good results in speed and performance,and it can be optimized in parallel.Integrating the advantages and disadvantages of GPU,ASIC,and FPGA in terms of data processing methods,speed,power consumption,and price,the dual-core ARM+FPGA heterogeneous platform Zynq was finally selected for deployment research.The main research contents are as follows:1.On the basis of the Yolo V4 network model,a Yolo V4-sim model is proposed according to the characteristics of the computing and storage resources of the embedded platform.Weighing the calculation delay and loss of precision,quantize the 32-bit floating-point data in the deep learning algorithm into 8-bit integers,and optimize the normalization method of the entire model,and propose BGN(Batch Group Normalization)to replace BN(Batch Normalization),Merge the three dimensions of channel,height and width into a new dimension,divide the new dimension into feature groups,and calculate the statistics of the entire small batch and feature groups.In the prediction frame extraction part,the boundary point enhancement extraction strategy is adopted to reduce the number of redundant frames and filter out harmful background information.2.According to the basic principles of convolutional neural network hardware acceleration,it is proposed to use hierarchical convolution and point convolution to simplify the calculation complexity,and propose the idea of data bypass to improve accuracy,and on the basis of 32bit-8bit quantitative reasoning algorithm,Match the characteristics of the embedded INT16 register to set the multiply-accumulate anti-overflow condition constraint.Although the average detection accuracy m AP is reduced by 5%,the detection speed is increased by about 6 times.3.According to the basic principles of convolutional neural network hardware acceleration,it is proposed to use hierarchical convolution and point convolution to simplify the calculation complexity,and propose the idea of data bypass to improve accuracy,and on the basis of 32bit-8bit quantitative reasoning algorithm match the characteristics of the embedded INT16 register to set the multiply-accumulate anti-overflow condition constraint.Although the detection accuracy m AP is reduced by5%,the detection speed is increased by about 6 times.In this paper,a portable target detection system based on deep learning is researched,weighing the influence of detection accuracy,algorithm complexity,power consumption,speed and other factors,and finally realizes the design of a low-power,low-cost,and high-real-time target detection system.The accuracy has reached 71.7%,and the target detection speed has reached 24 FPS,which provides a certain reference value for the future deployment of the target detection algorithm based on deep learning and the acceleration of its hardware implementation.